sgasser/podcast-mcp
If you are the rightful owner of podcast-mcp and would like to certify it and/or have it hosted online, please leave a comment on the right or send an email to dayong@mcphub.com.
Podcast MCP is a professional multi-speaker podcast generation tool integrated into Claude Desktop, utilizing high-quality AI voices and local processing for privacy.
Podcast MCP 🎙️
Generate professional multi-speaker podcasts directly inside Claude Desktop.
✨ Features
- Real AI Voices: Uses high-quality voices (not robotic text-to-speech).
- Multi-Speaker: Supports up to 4 distinct speakers automatically.
- Zero Config: Works instantly after installation.
- Privacy First: All processing happens locally on your machine.
🚀 Installation
Prerequisites
You need uv installed on your system via Homebrew (required for Claude Desktop to find it):
# macOS
brew install uv
# Linux/Other
# Use Homebrew on Linux, or ensure uv is in system PATH
# See: https://docs.astral.sh/uv/getting-started/installation/
Important: Installing via Homebrew ensures uv is in the system PATH that Claude Desktop can access.
Install the MCP Server
- Download
podcast-mcp.mcpbfrom the releases. - Drag and drop the file into Claude Desktop.
- That's it! 🎉
💡 Tips for Best Quality
- Speed: If it feels too slow, increase
TTS_SPEEDin.env(e.g. to 1.3 or 1.5). - Pauses: You can adjust
DEFAULT_PAUSE_MSin.envif the gaps between speakers are too long. - Remember: Restart Claude Desktop after changing
.envfor changes to take effect.
💡 How to Use
Simply ask Claude to create a podcast. You don't need to worry about formats or technical details.
Try this:
"Generate an audio podcast about the history of coffee."
Works in 20+ languages (German example):
"Erstelle eine Audio-Podcast-Datei über die Geschichte des Kaffees."
Claude will handle everything: writing the script, assigning voices, and generating the audio file.
⚙️ Configuration (Optional)
The default settings work great, but you can customize them if you want.
A .env file is automatically created on first run. To customize:
- Open the
podcast-mcpfolder (where Claude installed it). - Edit the
.envfile to customize:- Voices: Choose from 50+ available speakers.
- Speed: Adjust speaking rate (default is 1.2x).
- Output Folder: Change where files are saved (default:
~/Downloads). - Log Level: Set to DEBUG for detailed logs (default: INFO).
- Restart Claude Desktop for changes to take effect.
❓ FAQ
Where are my podcasts saved? By default, they appear in your Downloads folder.
Can I use my own voices? Currently, we support the built-in high-quality voices from Coqui XTTS-v2.
Does it work offline? Yes! The first time you run it, it downloads the AI model (~2GB). After that, it works entirely offline.
Where can I find debug logs?
Claude Desktop automatically captures all logs to ~/Library/Logs/Claude/mcp-server-Podcast MCP.log. You can view them with:
tail -f ~/Library/Logs/Claude/mcp-server-Podcast\ MCP.log
To change the log level, edit MCP_LOG_LEVEL in .env (options: DEBUG, INFO, WARNING, ERROR).
🙏 Acknowledgments
This project is built on excellent open source software:
- Coqui TTS - High-quality text-to-speech with XTTS-v2 model
- MCP - Model Context Protocol for Claude integration
- pydub - Audio processing and manipulation
- uv - Fast Python package manager
- python-dotenv - Environment configuration management
License: MIT | Author: Stefan Gasser